On-site lithology identification for tunnel face via self-supervised learning on TBM muck images

Reliable lithology identification in tunnel boring machine (TBM) construction is critical for optimizing operational parameters and reducing cutter wear. Existing approaches predominantly rely on supervised learning models, which face dual challenges of limited labeled data in the early stages of new projects and insufficient adaptability to dynamic data accumulation conditions. This study employs a self-supervised learning approach for lithology identification using muck images collected from TBM excavation. A simple framework for contrastive learning of visual representations (SimCLR)-based model is employed for self-supervised pretraining of lithological features, followed by downstream classification to assess performance. A muck image acquisition system was deployed at the Yinchao-jiliao 2-3 tunnel site, and a dataset containing 18 149 images across six lithology classes was constructed. To simulate increasing labeled data during excavation, different training-test ratios were designed to evaluate and compare the transferability and predictive performance, as well as update efficiency of the self-supervised SimCLR model (SSL-SimCLR) against the supervised ResNet50 model (SL-ResNet50). The results show that: (1) Under limited labels (20% training data), SSL-SimCLR achieved an F 1 -score of 0.88, outperforming SL-ResNet50 ( F 1 -score of 0.76) by 15.8%; (2) The SSL-SimCLR model exhibited a clear advantage when the training set ratio was below 50%, while both models achieved gradually comparable performance after the training data increased to 60% and ultimately reached an F 1 -score of 0.92; (3) In conditions requiring continuous model updating during excavation, SSL-SimCLR demonstrated significant deployment efficiency, with a total update time of 702.9 s, which accounts for only 3.8% of the 18 361.4 s required by the SL-ResNet50 for full-parameter retraining. These results indicate that the self-supervised approach delivers robust lithology recognition performance, particularly under limited training data, and offers a solution for rapid model updating in TBM tunnelling.